| Intelligent drive has a good development prospect in improving traffic safety,relieving traffic pressure and improving driving experience,and it has always been a hot topic of research.In the intelligent driving system,a very important part is to detect the targets on the road.At present,the mainstream target detection algorithms,such as YOLO and R-CNN series,have been the focus of research because of their high accuracy.However,these algorithms have high requirements for equipment because of the large amount of parameters.If they are directly used in vehicle-mounted equipment,it will cost a lot of money.Therefore,it is necessary to study a lightweight algorithm that can be applied in vehicle-mounted systems to reduce the cost and achieve popularization.The research results of this paper are as followss:1.The principle of object detection based on deep learning is deeply studied,and a road object detection algorithm based on improved feature extraction network is realized.based on YOLOv4 network,the separable convolution module Mobilenetv2 is used to replace the feature extraction part of YOLOv4 network,which greatly reduces the overall parameters of the network from 64.4M to 12.7M.The experimental results show that the detection speed of the improved algorithm has been improved,and the average detection time of each image has been reduced from the original 43 ms to 32 ms,which reduces the amount of network computation and reduces the requirements of on-board computing equipment.However,after the improvement,the detection accuracy of the network decreases slightly,and because the movement of the target and the occlusion between the targets will cause the problem of missed detection,the target detection algorithm needs to be further improved.2.In order to solve the problem of missing detection in the process of target detection and improve the accuracy of detection,this paper implements a road target detection algorithm combining target tracker,which uses the method of target detection and target tracking,takes the target tracking based on Kalman filter as an auxiliary means,combines the target motion relationship between frames in the video sequence,updates the detected objects by matching detection with prediction,and keeps the original motion trajectory to predict the next state as an edge frame to reduce missing.The experimental results show that the detection accuracy of the detection algorithm for each category has been improved after the fusion of the target tracker,and the m AP value has increased from 70.20% to 76.04%.3.In order to solve the problem of tracking errors when multiple targets are close to each other,this paper improves the target tracking algorithm,and on this basis,proposes a road target detection algorithm based on the improved target tracker,which uses the IOU of the detection frame and tracker prediction frame and the color histogram distance as a new cost function to match.The experimental results show that,compared with the target tracking algorithm that only uses IOU as the cost function,the number of id information switching times of the improved algorithm is reduced from 415 to 258,which improves the stability of the tracking process and proves the feasibility of the algorithm. |
